MiniMax H3 Trained 2x Latent Upscaler Ships With ComfyUI Nodes

Tr1dae released a trained clean-latent 2× upscaler for MiniMax H3, packaged as the ComfyUI-MiniMaxH3_LatentUpscaler custom node pack with a companion checkpoint on Hugging Face (h3_clean_latent_upscaler_film_epoch200.safetensors). The node hard-locks upscaling to exactly 2× and runs entirely in latent space, so a workflow can go from a low-resolution first pass to a high-resolution second pass without leaving the sampler.

What it is

Unlike a conventional upscaler that sharpens a finished render, this tool doubles the spatial latent dimensions between MiniMax H3 samplers: render small and cheap first, upscale the clean latent 2×, then continue denoising at full resolution.

• Doubles only the spatial latent dims; temporal length is unchanged.
• Uses a trained 2× network by default (method = learned model), learned as a correction over bilinear latent interpolation against pixel-upscaled/re-encoded teacher latents, with decoder-aware, SSIM, spatial and temporal losses.
• Preserves the audio stream unchanged. audio_denoise set to 0 locks pass-1 audio entirely; 1 lets pass 2 fully remix it. Values in between are not supported.
• Performs NestedTensor-aware CONST re-noise at sigmas[0] — stock LatentUpscaleBy / AddNoise break on MiniMax's joint audio-video latents (video [B,24,T,H/16,W/16] + audio [B,32,2,T_audio]).

The pack ships several nodes: MiniMax H3 Latent Upscale Combined (the main apply node), norm-preserving Latent Blur / Sharpen / Contrast helpers, and a Save / Load Package + Upscale Collect system that stashes pass-1 latents and conditioning for deferred high-quality upscaling with an HQ timeline and MP4 / FCP7 XML export.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler

The trained checkpoint auto-downloads on the first learned model run (SHA-256 verified) into ComfyUI/models/h3_latent_upscalers/. The architecture code is loaded from the sibling ComfyUI-H3-Latent-Upscaler-Mamad8 install, which is also referenced by the earlier 2× latent upscaler news.

Community results

The Banodoco community has been benchmarking it since mid-August. foxydits reports his Seed Hunter workflow (built on this node) renders a 0.5 MP base up to 1080p in under 5 minutes on an RTX 3090 (discord). DiXiao combined it with ComfyUI-MiniMax-H3-SPEED and the Turbo LoRA, producing a 10-second 1344×768 video in about 150-180 seconds, calling it "better than the standard upscaling method the node used and almost as fast" (discord).

A comparison from foxydits' early Seed Hunter test shows the upscaled pass against the 0.2 MP original:

Latent-upscaled 704×1216	Original 0.2 MP render

Context

The release fills the gap left by MiniMax's own H3 Regenerate-2K upscaler, which the team said in an AMA is still being optimized for efficiency before open-sourcing. In the meantime community latent upscalers are the practical answer, and the Tr1dae pack adds a trained-network option on top of the earlier interpolation-based approaches.